A system and method for energy storage battery state simulation and prediction based on digital twinning

Abstract

The application provides a kind of energy storage battery state simulation and prediction system and method based on digital twinning, aiming at solving the problems of inaccurate state modeling, insufficient perception granularity, uninterpretable prediction and response lag of existing energy storage system.The method uses equivalent thermal model, electrochemical model and aging mechanism model to model the battery, introduces adaptive gating parameter fusion mechanism and data characteristics, realizes cross-scale state mapping and feature reconstruction from cell to system level;Using the electric-thermal-aging coupling relationship to generate baseline trajectory, combined with the data-driven residual correction mechanism to realize multi-time domain closed-loop prediction, identify potential faults and risks in advance through multi-time domain anomaly judgment system;Build a three-dimensional topological model to dynamically display state parameters, generate hierarchical alarm signals and provide operation suggestions.The system constructed by the application has high precision, interpretability, prediction and visualization capabilities, which can significantly improve the intelligent operation and maintenance level of energy storage battery system.

Description

Technical Field

[0001] This invention belongs to the field of lithium batteries for energy storage systems, and specifically relates to a system and method for simulating and predicting the state of energy storage batteries based on digital twins. Background Technology

[0002] With the development of new energy, lithium batteries, as energy storage units, are widely used in power systems, microgrids, industrial and commercial energy storage and other fields. In order to ensure the stable operation and safe maintenance of energy storage systems, battery system status monitoring, health assessment and operation trend prediction have become important research directions in the industry.

[0003] Currently, due to limitations such as real-time requirements, embedded hardware computing power, sensor costs, and system integration complexity, engineering practices typically employ modeling methods with simple structures and low computational overhead, such as equivalent circuit models (ECM), thermal models, or empirical rule models. These models have good engineering feasibility and can achieve SOC / SOH estimation and temperature change simulation under normal operating conditions. However, in practical applications, energy storage batteries face nonlinear problems such as variable loads, complex environments, and unpredictable aging paths. Traditional models struggle to accurately characterize their dynamic behavior, leading to significant simulation bias and prediction uncertainty.

[0004] To improve predictive capabilities, some research and systems have attempted to introduce data-driven machine learning methods, such as neural networks, random forests, and LSTM algorithms, to predict battery life or abnormal states. This has improved prediction accuracy to some extent. However, existing technologies still have significant shortcomings in terms of modeling accuracy, state awareness granularity, prediction interpretability, and visualization depth, which hinders the development of energy storage battery systems towards higher levels of intelligence, autonomy, and safety. This is the limitation of existing technologies.

[0005] In view of this, the technical solution of the present invention provides a high-precision, interpretable, predictive and visualization-enabled digital twin system for energy storage batteries, which is necessary to solve the above-mentioned defects in the prior art. Summary of the Invention

[0006] This invention addresses the problems of inaccurate state modeling, insufficient sensing granularity, uninterpretable predictions, and lag response in existing energy storage systems. It proposes a method and system for simulating and predicting the state of energy storage batteries based on digital twins. By integrating physical mechanism models and data-driven methods, a multi-level and multi-dimensional virtual mapping body is established to achieve real-time sensing, dynamic simulation, and trend prediction of battery operating status, thereby solving the aforementioned technical problems.

[0007] The technical solution of this invention is as follows:

[0008] In a first aspect, the present invention provides a method for simulating and predicting the state of energy storage batteries based on digital twins, specifically including the following steps:

[0009] Step S1, the step of constructing a multi-scale digital twin model, in which:

[0010] Multiphysics modeling of the battery system is performed using electrical, thermal, and aging sub-models; a dual-frequency strategy and extended Kalman filtering are employed to achieve dynamic updating of model parameters; adaptive gating parameters are introduced to dynamically weight and fuse the outputs of the physical model and the data-driven model, combined with a residual compensation mechanism to achieve high-precision state estimation; a multi-level structure mapping engine is constructed to achieve bidirectional mapping between cells, packs, clusters, and modules; and a multi-scale digital twin model is built.

[0011] Step S2, the trend simulation and state prediction step, in which:

[0012] Based on step S1, a coupled state mechanism of electro-thermal-aging is constructed. Through cross-scale state convergence and feature reconstruction, multi-level state co-evolution is achieved. A predictive baseline for state evolution is generated through a physical model. A data-driven model is used to correct physical prediction biases. A multi-time-domain anomaly judgment system is constructed. The physical residuals and data residuals of key monitoring indicators are calculated. Combined with rate of change detection, time-series anomaly detection models, and long-term trend analysis, anomaly localization is achieved through an electro-thermal-aging collaborative feedback mechanism. Through feature contribution quantification, counterfactual simulation, and feature attribution analysis, combined with attention mechanisms and causal analysis, multi-level explanatory information is generated.

[0013] Step S3, the visualization and intelligent early warning step, in which:

[0014] A three-dimensional spatial topology model of cell-pack-cluster-module is established. The cell status is aggregated to the PACK level, cluster level, and module level through hierarchical mapping rules, maintaining the topological relationship in physical space. Key status parameters are dynamically bound to visual primitives, and the values ​​are mapped to visual attributes through normalization, weighting, and topological constraints. Key anomaly indicators are recorded and visualized in time series. Alarm signals are generated based on trend prediction results and anomaly trajectories and are processed in a hierarchical manner, using adaptive threshold calculation, multi-dimensional risk scoring, and historical pattern comparison.

[0015] The steps for constructing the multi-scale digital twin model further include the following steps:

[0016] S101, Steps for physical mechanism modeling, in which:

[0017] Equivalent thermal model, electrochemical model and aging mechanism model are used to simulate the response behavior of battery under different dimensions. Through electrical sub-model, thermal sub-model and aging sub-model, the current-voltage dynamic response, heat conduction and heat dissipation process and capacity decay and internal resistance growth law of battery are modeled respectively.

[0018] Furthermore, the electrical sub-model adopts a second-order Thevenin equivalent model to dynamically model the cell terminal voltage and polarization voltage.

[0019] Furthermore, the inputs to the electrical sub-model include real-time current, voltage, ambient temperature, initial SOC, and initial voltage polarization state; model parameters, including ohmic internal resistance, electrochemical polarization resistance, and capacitance, are obtained through experimental calibration or historical data fitting.

[0020] Furthermore, during the operation of the electrical sub-model, a dual-frequency strategy is adopted to calculate and update the parameters to obtain output parameters, including real-time SOC, terminal voltage prediction, and polarization voltage components. The output here not only provides the basis for heating calculation for the subsequent thermal sub-model, but also provides electrical state reference for the aging sub-model and the data-driven model, supporting multi-physical quantity coupling analysis and cell-level performance prediction.

[0021] Furthermore, in the dual-frequency strategy, the high frequency indicates that the model parameters remain unchanged in the short term. The current parameters are used to simulate the polarization effect inside the cell in real time through a second-order RC model to calculate parameters such as voltage polarization component, terminal voltage, and SOC. The low frequency indicates that under fixed period or triggering conditions, the historical parameters and operational observation data are used to correct the model parameters through extended Kalman filtering to obtain updated impedance, capacitance, and polarization state parameters.

[0022] Furthermore, the thermal sub-model adopts a hierarchical thermal network structure, in which the cell level adopts a multi-node thermal model, and the PACK, cluster, and compartment system levels adopt equivalent thermal node models to simulate the thermal coupling and temperature distribution between cells, PACK, clusters and compartments.

[0023] Furthermore, the inputs to the thermal sub-model are the output voltage, SOC, real-time current, ambient temperature, and cooling medium state of the electrical sub-model; model parameters, including cell specific heat capacity, thermal conductivity, and convective heat transfer coefficient, are obtained through experimental measurement or calibration.

[0024] Furthermore, in the process of processing, each cell is abstracted into several hot nodes, each node representing a physical region of the cell; the node parameters are assigned according to the physical location, and the temperature of each node of the cell is calculated using the energy conservation and heat conduction equations. Taking into account the Joule heat generated by the current, the heat conduction between adjacent cells and external heat dissipation, the hot spot locations of the temperature distribution between the cell and the compartment are obtained.

[0025] Furthermore, the Joule heating formula is as follows:

[0026]

[0027] in, It is Joule fever. For current, For resistance;

[0028] Furthermore, the node parameters include specific heat, thermal conductivity, and convective heat transfer area, and the physical region includes key heat capacity units such as positive electrode plates, negative electrode plates, electrolyte, or shell.

[0029] Furthermore, the output of the thermal sub-model is cell temperature and hotspot information, which provides temperature input for the aging sub-model and can be used for twin visualization mapping.

[0030] Furthermore, at the PACK, cluster, and cabin levels, an equivalent single-node thermal model of the battery cell is used for modeling. The next-level structure is thermally equivalent to a lumped parameter node, used to describe the overall average temperature state and heat exchange behavior of that level. The heat capacity of the equivalent single node is obtained by weighting the heat capacities of its subordinate units by quantity or mass, and the equivalent thermal resistance is determined by the structural connection relationship, cooling path, and installation method.

[0031] Furthermore, the aging sub-model models cell capacity decay and SOH evolution based on empirical or semi-physical equations.

[0032] Furthermore, the inputs to the aging sub-model include cell SOC, temperature, charge-discharge cycle count, and historical capacity decay data. The model parameters, including the aging rate constant and decay coefficient, are obtained through experimental fitting.

[0033] Furthermore, the aging sub-model calculates capacity loss based on conditions such as SOC, temperature, and cycle count, and predicts the remaining capacity and SOH evolution trend of the cell. The specific expression is as follows:

[0034]

[0035] in, This indicates the degree of degradation of the battery's health status per unit time. The larger the value, the faster the battery ages. This indicates the battery's state of charge, a key parameter affecting aging. Representing temperature and charge / discharge cycle count, respectively, the function... It can be implemented by including a temperature Arrhenius term and a load / cycle term, so that the aging rate increases with increasing temperature, cumulative cycles, and enhanced high load conditions;

[0036] Furthermore, the output of the aging sub-model includes the cell's remaining capacity, SOH, decay rate, and capacity inconsistency index, providing a physical basis for trend prediction, anomaly identification, and multi-scale twin mapping.

[0037] Furthermore, the capacity inconsistency index is calculated from the SOH distribution of the cell group, including range, standard deviation or quantile difference;

[0038] Furthermore, to improve the online accuracy and adaptability of the physical model, the physical mechanism introduces an extended Kalman filter (EKF) to update the cell state at high and low frequencies. The inputs of the EKF include the predicted states output by the electrical, thermal, and aging sub-models, as well as real-time measurements. The state estimation increment is calculated through the state equation and the measurement equation.

[0039] Furthermore, the EKF employs a dual-timescale update strategy: high-frequency updates are used to perform real-time corrections on fast states such as SOC, polarization components, and temperature to quickly track operational fluctuations; low-frequency updates are used to recursively estimate slowly changing parameters, including polarization resistance / capacitance, equivalent thermal parameters, and aging rate parameters, enabling the model to self-correct as operating conditions and life stages change.

[0040] Furthermore, the predicted state includes SOC, terminal voltage, polarization voltage, temperature, and capacity loss, and the real-time measured values ​​include current, voltage, and ambient temperature.

[0041] Furthermore, the model parameters include polarization resistance, capacitance, thermal conductivity, and aging rate;

[0042] Furthermore, the EKF outputs the corrected cell-level SOC, voltage polarization component, temperature, and SOH, and generates derived features such as capacity imbalance and hot spots, providing a reliable physical benchmark for subsequent data-driven compensation, anomaly prediction, and visualization.

[0043] Step S102, the physical-data fusion modeling step, in which:

[0044] By introducing adaptive gating parameters constrained by physical consistency into the model structure, and integrating the mechanistic and data characteristics of cell operation, the problem of insufficient modeling accuracy and difficulty in characterizing nonlinear features under complex operating conditions of the physical model is solved. This balances physical consistency and nonlinear expressive ability, thereby achieving accurate estimation of key state quantities such as SOC, SOH, terminal voltage and temperature under complex operating conditions, and providing a stable and reliable state benchmark for subsequent trend prediction, anomaly identification and multi-level twin mapping.

[0045] Furthermore, the inputs to the steps include cell SOC, terminal voltage, polarization voltage, temperature, and SOH obtained from physical mechanism modeling, as well as historical operating data, including charge and discharge current, voltage, temperature, depth of charge and discharge (DOD), rate, number of cycles, and operating condition markers, etc.

[0046] Furthermore, a baseline state equation is constructed based on the electrical sub-model, thermal sub-model, and aging sub-model to characterize the dynamic evolution of the cell under the influence of current, temperature, and aging. The baseline state equation serves as the physical constraint and prediction reference for the fusion model, and is used to limit the output space of the data-driven model.

[0047] Furthermore, after the running data is interpolated, normalized, and anomaly detected, multidimensional time series features are formed based on the time window division, and derived indicators such as voltage change rate, temperature gradient, and capacity decay rate are constructed to enhance the model's ability to describe nonlinear features.

[0048] Furthermore, in the data-driven stage, a Long Short-Term Memory (LSTM) network is used to capture the dynamic dependence and cross-time characteristic coupling of the cell during the cycling process. Through a gating mechanism, the LSTM is used to represent complex nonlinear relationships such as SOC-voltage and temperature-degradation.

[0049] Furthermore, during the training process, long-term dependency information is stored and updated through a time-step expansion gating mechanism. The loss function is optimized primarily by the mean square error of the predicted terminal voltage, SOC, and temperature. Capacity inconsistency index and temperature gradient features are introduced as auxiliary constraints to improve the fitting ability to boundary conditions and atypical modes.

[0050] Furthermore, after the model training is completed, a short-term prediction of the future state of the battery cell is generated based on the current state sequence.

[0051] Furthermore, in the model fusion stage, a gating function g(t) is introduced to control the physical model prediction and data-driven prediction. Dynamic weighted fusion is achieved by adaptively adjusting model weights. The gate function takes the current SOC, temperature, aging degree, and predicted residual characteristics of the battery cell as input. It dynamically calculates the weight allocation relationship between the physical model and the data-driven model through nonlinear mapping, thereby realizing adaptive switching and collaborative output of the two types of models at different operating stages. The fusion form is as follows:

[0052]

[0053] Among them, g(t) is a gate parameter, which is obtained by nonlinear mapping calculation based on the real-time state of the battery cell. It can dynamically adjust the contribution ratio of the two types of models according to different operating conditions. The predicted values ​​are from the data-driven model. These are the predicted values ​​from the physical model; This is the final fusion prediction value;

[0054] Furthermore, when the system is in a steady state or linear range, the physical model dominates; while in the stages of strong nonlinearity, abrupt feature changes, or aging, the data-driven model is used for correction and compensation, thereby achieving a dual guarantee of prediction accuracy and physical consistency.

[0055] Furthermore, to reduce model bias and improve long-term operational stability, the system introduces a residual compensation mechanism: by calculating the difference between the physical prediction and the fused prediction and applying a dynamic correction coefficient α(t), the fusion result is corrected a second time, as shown in the following expression:

[0056]

[0057] in, This is a dynamic correction factor; It is the final accurate prediction result output by the system. This is to connect with the intermediate prediction results mentioned earlier.

[0058] Furthermore, the final output cell's core state parameters at a given moment, such as SOC, terminal voltage, polarization voltage, temperature, and SOH, are generated simultaneously, along with derived characteristics such as capacity inconsistency, hotspot indicators, and degradation rate. These outputs serve as standard inputs for the multi-level twin system, providing accurate and interpretable foundational support for multi-level twin mapping, trend prediction, and anomaly identification.

[0059] Step S103, the steps of the multi-level structure mapping engine, specifically also include:

[0060] A multi-level structure mapping engine for energy storage systems is constructed to realize cross-scale state mapping and feature reconstruction of cells-packs-clusters-modules in energy storage systems. The mapping engine takes the cell-level fused state as the basic input and establishes a continuous association from the micro-cell state to the macro-system performance by introducing electrical topology constraints, thermal coupling relationships and physical attribute weights, thereby achieving data consistency, physical constraint consistency and feature traceability.

[0061] Furthermore, the multi-level structure mapping engine adopts a bidirectional mapping mechanism of "uplink aggregation - downlink reconstruction". Uplink aggregation is used to aggregate the cell-level state to the PACK, cluster and module levels step by step. Downlink reconstruction is used to reverse calculate the cell-level state based on historical mapping relationships and weight distribution when system-level analysis or anomaly localization is required, thereby realizing bidirectional state synchronization across scales and providing a unified state expression basis for digital twin systems.

[0062] Furthermore, the inputs to the multi-level structure mapping engine include key state variables output from the cell-level fusion model, as well as derived features generated by the data-driven module, while also incorporating environmental layer information to enhance the integrity and physical accuracy of cross-level mapping.

[0063] Furthermore, the key state quantities include SOC, terminal voltage, polarization voltage, temperature, and SOH; the derived characteristics include capacity inconsistency, temperature difference hotspots, aging rate, etc.; and the environmental layer information includes liquid cooler unit status, ambient temperature, air-cooled / liquid-cooled medium parameters, etc.

[0064] Furthermore, the input data is standardized, dimensionally normalized, and weighted. An adaptive weighting mechanism based on physical attributes and spatial location is introduced. The weighting coefficients can be adaptively adjusted according to the importance of the cells, their location distribution, and their historical health status, so that the characteristics of different cells have a unified dimension and numerical range, ensuring numerical comparability and computational stability during multi-level aggregation, so as to reflect the overall performance.

[0065] Furthermore, during the uplink aggregation stage, the multi-level structure mapping engine aggregates cell-level data into PACK-level, cluster-level, and compartment-level states based on the electrical and thermal topology of the battery pack, including series and parallel relationships, physical spatial location, and thermal coupling paths.

[0066] The convergence process employs a multi-parameter weighting strategy, during which:

[0067] Voltage aggregation follows the series and parallel circuit rules, combining layer by layer through the topology matrix to obtain the total voltage of the PACK, cluster, and compartment. SOC aggregation adopts a capacity-weighted average method, with the PACK-level SOC calculated using the effective capacity of individual cells as the weight, and the cluster and compartment-level SOC iteratively aggregated using the rated capacity of the PACK and cluster levels as the weight. Temperature aggregation is based on a heat capacity-weighted average, and simultaneously calculates consistency indicators such as maximum temperature difference and standard deviation to reflect the degree of thermal imbalance within the PACK, cluster, and compartment. When a State of Health (SOH) aggregation is performed, a capacity-weighted strategy or a conservative worst-case strategy can be adopted according to application requirements to achieve different orientations for performance characterization or safety characterization.

[0068] Furthermore, to enhance the model's engineering robustness, the mapping engine simultaneously calculates consistency metrics for each level during the convergence phase, including voltage standard deviation, SOC dispersion, and temperature range. Based on these consistency metrics, it establishes inconsistency criteria between levels to detect potential individual cell failures or system anomalies. If missing or abnormal cell data is detected, the engine completes the state through a multi-source feature reconstruction algorithm.

[0069] Furthermore, the completion process can comprehensively adopt the following methods: interpolation correction based on the thermal coupling relationship of adjacent cells; dynamic compensation based on time series models; and weighted estimation based on the historical distribution pattern of cells, thereby achieving data continuity and spatial integrity without increasing the cost of additional sensors.

[0070] Furthermore, during the downlink reconstruction phase, when the system needs to perform detailed analysis or local state deduction at the PACK or cell level, the mapping engine can reverse calculate the cell-level state based on the historical mapping matrix and weight distribution, supporting single-unit hotspot backtracking, health state reconstruction and local anomaly tracking, and realizing bidirectional state mapping across the entire link.

[0071] Furthermore, the output includes the cell-level status after downlink reconstruction, as well as the PACK-level, cluster-level, and module-level status after uplink aggregation, specifically covering the SOC, voltage, temperature, aging status, and derived abnormal indicators at each level; at the same time, it records the maximum value of the status parameters at each level and hotspot nodes, providing standardized input for subsequent trend prediction, anomaly identification, and visualization.

[0072] The beneficial effects of step S1:

[0073] To address the technical challenge of traditional modeling methods failing to simultaneously consider the nonlinearity, multi-physics coupling characteristics, and fine-grained state mapping of battery systems, thus hindering the construction of high-precision, real-time responsive virtual mapping entities, the following solutions are proposed: A hybrid modeling method based on multi-source fusion is introduced. This method compensates for nonlinear relationships such as SOC-voltage and temperature-aging by introducing a data-driven model under physical model constraints. Adaptive gating parameters are used to dynamically adjust the weights of the two types of models, effectively compensating for modeling errors in pure physical models under conditions of strong nonlinearity and significant aging. This approach balances physical constraints and data learning, improving simulation accuracy and dynamic adaptability. A unified virtual mapping of multi-granularity and multi-dimensional states is achieved, supporting fine-grained state reproduction from cell to system level. This ensures physical consistency and data continuity in state evolution across different levels, providing a reusable and scalable foundational model for multi-level digital twin mapping. Furthermore, an online model calibration and lifecycle evolution adaptive update mechanism is supported. Through residual compensation and dynamic parameter update mechanisms, the model can continuously correct prediction deviations based on real-time data during operation, possessing long-term evolution capabilities.

[0074] Step S2 specifically also includes:

[0075] Step S201: A step in the multi-field coupled prediction algorithm, in which:

[0076] This paper utilizes the complex coupling relationship between physical, thermal, and electrochemical aging to perform multi-time-domain prediction. A multi-physics coupling prediction algorithm is employed to simulate and predict the state evolution of energy storage batteries in future multi-time domains. The algorithm uses the cell-level state output in step S1 as the initial condition and simultaneously describes the coupling feedback relationship between electrical dynamics, thermal transfer, and aging evolution within a unified state space. This addresses the problems of unstable predictions and unreliable long-term trends under complex operating conditions caused by single models. The algorithm's input includes the cell-level state vector, historical observation sequences, and future operating condition assumptions (load plan, temperature control strategy, ambient temperature boundary, etc.). The output includes the predicted trajectory, confidence interval, and risk metric for multiple levels from cell to PACK to cluster to compartment.

[0077] Furthermore, the dynamic evolution of the cell terminal voltage and polarization component as a function of current fluctuation is calculated using an electrical sub-model. At the same time, the thermal sub-model is combined to simulate the heat conduction, convection heat transfer and external heat dissipation effects of the cell, PACK, cluster, and inter-cell to obtain the temperature distribution.

[0078] Furthermore, the aging sub-model updates SOH and related aging indices based on slow variables such as temperature and cyclic load dynamics, and uses the aging state feedback to correct the electrical and thermal model parameters, forming a closed-loop coupling mechanism to achieve real-time interaction and dynamic adjustment between different physical quantities. It maps the change in SOH / internal resistance to the drift of ohmic internal resistance and polarization impedance, and introduces the change in heat exchange efficiency or thermal resistance into the update of thermal model parameters, thereby forming a closed-loop coupling link of "electric drive - thermal response - aging accumulation - parameter drift - electrothermal re-response", so that the mutual influence between different physical quantities can be continuously reflected in the prediction process.

[0079] Furthermore, in implementation, the battery cell is the smallest computational unit, and the state vector of each cell includes SOC, terminal voltage, polarization components, temperature, and SOH. A uniform time step is used for prediction. The state equations are iteratively updated, forming a multi-step forward propagation process: In each time step, the terminal voltage and polarization components are first updated by the electrical model based on the given operating conditions (current, cooling state, environmental boundary), then the temperature state is updated by the thermal model based on the heat generation power and thermal network, and finally the SOH / internal resistance drift is updated by the aging model based on the temperature and cumulative load, and this drift is written back to the electrical and thermal parameters of the next step, realizing the dynamic self-consistent update of the coupling parameters. To ensure cross-scale consistency, the PACK, cluster, and compartment-level states are obtained in real time from the cell-level states through topological constraints and weighted rules, and are used to output system-level prediction results and consistency indicators, ensuring cross-scale data consistency.

[0080] Furthermore, the generation of the physical prior trajectory is achieved through multi-step forward propagation; the current cell state is used as the initial condition, and the discretized electro-thermal-aging coupling equation is iteratively calculated at a fixed time step to obtain the short-term prediction results, forming a physical baseline;

[0081] Furthermore, in short-term forecasts at the minute to hour level, high-frequency sampling data drives the dynamic updating of the physical model to ensure real-time responses to terminal voltage, SOC, and temperature. In long-term forecasts at the day to week and month to year level, since high-frequency measurements are unavailable, the system extends the physical baseline trajectory through historical cyclic statistics, typical load sequences, and low-frequency parameter corrections (such as temperature-dependent aging rates and internal resistance growth curves), making the long-term forecasts trend-based and providing a basic reference for residual learning.

[0082] Furthermore, historical observation sequences are combined with physical baseline residuals to capture nonlinear behaviors that physical models cannot fully reflect. The residual sequences are encoded in time series networks (such as LSTM, GRU, or lightweight Transformer) and combined with historical state features, including current change rate, SOC gradient, temperature distribution features, and aging accumulation, to learn systematic biases under electro-thermal-aging coupling.

[0083] Furthermore, the loss function during network training not only includes the mean square error of terminal voltage, SOC, and temperature, but also introduces physical consistency constraints (charge conservation, SOH monotonicity) and uncertainty constraints (NLL or quantum loss) to ensure that residual predictions can capture boundary conditions and atypical patterns while conforming to physical laws. In the prediction stage, the output residual sequence is combined with the physical baseline to generate multi-time-domain prediction results, achieving a closed-loop fusion of "physical prior + data residuals".

[0084] Furthermore, the fusion of multi-dimensional states adopts a unified cell-level state vector representation. In each time step, current and voltage update cell polarization and terminal voltage, the thermal model updates the individual cell temperature and calculates heat conduction and dissipation of adjacent cells, and the aging model updates SOH and internal resistance based on temperature and load. The aging state is then fed back to correct electrical and thermal parameters, forming a closed-loop feedback across physical fields. In PACK-level, cluster-level, and compartment-level convergence, SOC uses a capacity-weighted average, temperature uses a thermal capacity-weighted average, and SOH can be either capacity-weighted or a worst-case strategy. Combined with topological constraints and position weights, cross-scale state consistency is ensured. When missing or abnormal cell data exists, a feature reconstruction algorithm is introduced to recover the state, ensuring a continuous and robust prediction process.

[0085] Furthermore, through the above process, continuous state simulation and trend prediction from short-term minute-level to long-term month-level levels are achieved, while outputting confidence intervals and risk metrics to support operation and maintenance decisions and preventive maintenance. The innovation of this method lies in constructing an electro-thermal-aging closed-loop coupling by combining physical priors and residual correction. At the same time, it utilizes multi-time-domain residual learning and cross-scale state convergence to achieve full-process virtual simulation and trend prediction. This ensures that the prediction results maintain physical rationality while possessing nonlinear adaptability and engineering interpretability, providing a reliable foundation for subsequent abnormal state prediction and intelligent early warning.

[0086] Step S202, the steps of the abnormal state prediction mechanism, in which:

[0087] The aim is to identify potential hidden faults and risks in energy storage battery systems in advance, and to address the shortcomings of traditional prediction methods in anomaly detection and hidden fault identification due to their reliance on thresholds or single models.

[0088] Furthermore, based on the SOC, terminal voltage, polarization component, temperature, SOH and derived indicators output in step S1, and combined with historical observation data and future operating condition assumptions, a multi-time-domain anomaly judgment system is constructed to realize a multi-time-domain anomaly prediction and graded risk quantification system from individual cells to PACKs and even clusters and modules.

[0089] Furthermore, to fully capture potential abnormal behavior, anomaly information is first modeled. Based on the fused predicted SOC, terminal voltage, polarization components, temperature, SOH, and their derived characteristics, a key monitoring index system is constructed. These indicators include at least: cell terminal voltage and group dispersion, single-cell equivalent internal resistance or voltage-SOC offset rate, temperature difference and temperature rise rate between single cell and PACK / cluster, and aging rate deviation. For each index, physical residuals (observed value minus physical baseline predicted value) and data residuals (observed value minus fused predicted value) are calculated to reveal deviations that the physical model and data model cannot fully explain. The residual sequence forms a multi-scale signal in the time dimension. Anomaly characteristics are identified by analyzing its instantaneous changes, cumulative trends, and pattern evolution to characterize abnormal deviation behaviors that are difficult for both the physical and data models to explain.

[0090] Furthermore, through residual sequence analysis, multi-time-domain anomaly identification is performed at different time scales, including short-term, medium-term, and long-term. Short-term anomaly identification focuses on sudden events at the minute to hour level, employing rate of change detection and statistical thresholding methods, including exponentially weighted moving average (EWMA), cumulative sum (CUSUM), and Bayesian online change point detection. These algorithms calculate the residual change rate in real time, capturing rapid abnormal fluctuations in voltage or temperature to achieve highly sensitive alarms.

[0091] Furthermore, the mid-term anomaly identification focuses on the development of inconsistencies at the day-week level. It adopts a time-series anomaly detection model, uses LSTM-Autoencoder to reconstruct the residual sequence and calculate the reconstruction error, and IsolationForest to score the anomalies of the multidimensional state, thereby discovering the gradually accumulating deviations or anomaly patterns.

[0092] Furthermore, long-term anomaly identification targets potential failure risks from month to year. Based on RUL estimation, Bayesian survival analysis, and trend regression, it integrates cell aging status, temperature history, and capacity decay rate to predict possible failure time windows and changes in healthy life.

[0093] Furthermore, in terms of multi-physics collaborative analysis, the anomaly prediction mechanism maintains consistency with the electro-thermal-aging closed-loop coupling model in step S201. Polarization changes caused by current and voltage disturbances, the amplification effect of temperature distribution on local aging rates, and the reverse correction of electrical and thermal parameters by aging states are all incorporated into the anomaly determination process, enabling anomaly identification not only based on apparent data deviations but also reflecting the inherent imbalance mechanism under multi-physics coupling.

[0094] Furthermore, multi-dimensional information such as voltage, current, temperature, and SOH are mapped into a unified state vector, which is dynamically updated iteratively over time steps. Combined with topological constraints, anomaly detection results are located at the specific cell, PACK, cluster, and compartment levels, achieving graded risk quantification. The anomaly detection threshold and model parameters adopt an adaptive update strategy, automatically adjusting based on historical normal segment statistics or rolling quantiles, enabling the system to adapt to changes in operating conditions and data drift, while reducing the false alarm rate.

[0095] Furthermore, the output of this step includes abnormal indicator values, risk levels, potentially affected individual units, packs or clusters, and estimated remaining response time, providing a quantitative basis for graded alarms and preventive operations and maintenance.

[0096] Furthermore, through multi-time-domain and multi-level anomaly identification and prediction, the system can provide actionable risk information before anomalies occur, enabling the operation and maintenance team to respond quickly to local anomalies in the short term, and formulate maintenance strategies or adjust load / temperature control schemes in the medium and long term planning to ensure the safe and stable operation of the energy storage system.

[0097] Step S203, the step of the interpretability enhancement mechanism, in which:

[0098] The interpretability enhancement mechanism aims to transform the output of the fusion prediction model from a black-box result into traceable, quantifiable, and operationally guiding explanatory information, addressing the problem that existing data-driven predictions struggle to explain the basis for predictions and cannot support risk assessment and intervention decisions. This mechanism constructs an interpretable closed loop throughout the entire prediction process, from input features to prediction results, and then to risk sources and intervention strategies.

[0099] Furthermore, this mechanism constructs a complete interpretable prediction closed loop through input processing, feature contribution quantification, counterfactual simulation, feature attribution, and fusion display. Its inputs include the cell-level states (SOC, SOH, terminal voltage, polarization voltage, temperature distribution, etc.) output by the fusion model in step S1, historical observation sequences, residual sequences, and future operating condition assumptions (load curves, temperature control strategies, charge and discharge rates, etc.), providing basic data for subsequent interpretable analysis.

[0100] Furthermore, in the input processing stage, the mechanism first normalizes, temporally windows, and performs hierarchical mapping on the input features to maintain the correspondence between each input variable in the time and spatial dimensions (cell / pack / cluster / module).

[0101] Furthermore, by utilizing the attention or gating mechanism within the time-series prediction model, the contribution weight of the input features at each time step to the final prediction result is calculated; the specific implementation method is as follows:

[0102] In LSTM, GRU, or lightweight Transformer networks, the influence of each input feature on the prediction target at different time steps is quantified into a numerical metric by backpropagating gradients to the hidden state and output, combined with the attention weight matrix. This contribution can be used to generate time-feature heatmaps or directly to evaluate key driving factors.

[0103] Furthermore, after the feature contribution quantification is completed, counterfactual simulation is introduced to evaluate the causal impact of a single factor on the predicted output. This is achieved using a digital twin model, where one or more input variables (such as future temperature, peak load, and discharge rate) are adjusted by a set amplitude, while keeping other conditions constant. The fused prediction model is then rerun to generate a new prediction trajectory. By comparing the difference between the original prediction and the counterfactual trajectory, the specific impact of the factor on SOC, SOH, temperature evolution, or RUL can be quantified, thus achieving causal interpretability.

[0104] Furthermore, feature attribution analysis is performed on the final fusion prediction results. Using approximation schemes such as SHAP or Integrated Gradients, the fusion prediction output is decomposed into the sum of the contributions of each input feature, including both the physical baseline and the residual correction. By performing feature attribution on the residual sequence, the relative contributions of factors such as current fluctuations, temperature changes, historical residuals, or aging conditions to prediction bias and risk scores can be clearly identified, providing quantitative evidence for operations and maintenance personnel.

[0105] Furthermore, in the output phase, the mechanism integrates attention / feature contribution, counterfactual simulation results, and attribution analysis to generate multi-level explanatory information, including time-feature level contribution heatmaps, causal sensitivity analysis, residual source decomposition, confidence intervals, and risk levels. By combining this with topological information, it can locate cells, PACK units, clusters, or modules where anomalies may occur and provide actionable intervention suggestions. Ultimately, this mechanism can provide complete, traceable, quantifiable, and cross-scale predictive explanations, explaining not only "what might happen in the future" but also clarifying "why it might happen" and "how to intervene," significantly improving the credibility and engineering deployability of the prediction results.

[0106] Beneficial effects of step S2:

[0107] To address the challenges of complex state changes in energy storage battery systems and the limitations of traditional methods in accurately predicting future states (insufficient prediction accuracy under complex operating conditions), the lack of interpretability in data-driven prediction processes, and the inability to provide reliable basis for operation and maintenance decisions, the following solutions are proposed: First, implement full-process virtual simulation and trend prediction of operating status and health parameters to overcome the shortcomings of traditional post-event response. Second, introduce a multi-physics collaborative prediction mechanism to enhance the predictive model's adaptability to degradation behavior under complex operating conditions. Third, adopt an interpretable modeling framework to clarify the role of SOC, temperature, aging status, and historical operating characteristics in the prediction results by quantifying the contribution of input features and decomposing residual sources in the fusion prediction model. This provides clear physical and data basis for the prediction conclusions, improving the credibility and engineering applicability of the prediction results.

[0108] Step S3 specifically also includes:

[0109] Step S301, the steps of the three-dimensional structure binding mechanism, in which:

[0110] A three-dimensional spatial topology model of cell-pack-cluster-module consistent with the actual energy storage system is established. Each cell, pack, cluster, and module corresponds to a unique entity object in the virtual three-dimensional environment, and its spatial coordinates, hierarchical membership, and topological connection relationships are maintained. The module input is the cell-level state (SOC, SOH, temperature, polarization component, capacity inconsistency, etc.) and derived characteristics output from steps S1 and S2.

[0111] Furthermore, through hierarchical mapping rules, the cell status is not only aggregated at the PACK, cluster, and module levels, but also maintains its topological relationship in physical space, enabling cross-scale synchronous state tracking. Each visualized object is fully mapped to the actual entity in terms of three-dimensional coordinates, hierarchical structure, and associated attributes. The visualized object, as a state-bearing node, includes at least SOC, SOH, temperature, terminal voltage, polarization components, and capacity inconsistency indicators. This solves the problem that traditional visualization remains at the macro level and cannot present the dynamic evolution of the cell, providing maintenance personnel with a fine-grained foundation for state awareness.

[0112] Step S302, the steps of the parameter dynamic mapping algorithm, in which:

[0113] In the parameter dynamic mapping algorithm, key state parameters (temperature, voltage, capacity decay rate, SOH, etc.) are dynamically bound to visual primitives to achieve intuitive feedback such as color gradients, geometric deformations, and animated flashing. Inputs include real-time state data, historical trajectory sequences, and user-defined mapping rules.

[0114] Furthermore, during the processing, numerical values ​​are mapped to visual attributes through normalization, weighting, and topological constraints, such as color gradients representing temperature, geometric deformation representing capacity decay, and flashing animations indicating abnormal fluctuations, thereby achieving synchronous and dynamic presentation of multidimensional data.

[0115] Furthermore, the algorithm not only supports the simultaneous display of multi-dimensional data, but also reflects the inconsistencies and hotspot locations between battery cells, thereby breaking through the limitations of traditional static display and achieving dynamic visualization with a high degree of consistency between state and structure.

[0116] Step S303, the steps of the abnormal trajectory evolution view, in which:

[0117] The system records and visualizes key anomaly indicators (voltage deviation, temperature rise rate, capacity loss rate, etc.) over time. The input is the state evolution data of each cell; during processing, the indicator trajectories are combined with 3D topology to achieve dynamic playback, trend curve overlay, and local highlighting, supporting source tracing analysis and anomaly evolution tracking.

[0118] Furthermore, this step solves the problem that traditional methods are difficult to analyze the development process of anomalies. Users can intuitively observe the evolution path of anomalies in time and space, support source tracing analysis and anomaly evolution tracking, and provide maintenance personnel with a panoramic view of anomalies from battery cells to PACKs, enabling early perception and analysis of potential risks.

[0119] Step S304, the intelligent early warning step, which generates an alarm signal based on the trend prediction results and abnormal trajectories of step S2 and performs hierarchical processing.

[0120] Furthermore, the intelligent early warning input includes predicted cell status, abnormal indicator trajectories, and an expert rule base. During processing, adaptive threshold calculation, multi-dimensional risk scoring, and historical pattern comparison are employed to provide early warnings for potential thermal runaway, cell inconsistencies, or capacity anomalies. The output includes tiered alarms, affected unit identifiers, and operational suggestions, with affected areas highlighted in a 3D interface. This mechanism solves the problem of slow response times in traditional manual intervention, achieving prediction-driven proactive early warning and enhancing operational efficiency and decision reliability through visualization.

[0121] The above methods overcome the limitations of traditional static displays and manual early warnings. Visual displays not only provide real-time status awareness, but also directly support prediction-driven operation and maintenance and risk management.

[0122] Secondly, the present invention provides a system for simulating and predicting the state of energy storage batteries based on digital twins, including a multi-scale digital twin model construction module, a trend simulation and state prediction module, and a visualization and intelligent early warning module.

[0123] The multi-scale digital twin model construction module, in which:

[0124] Multiphysics modeling of the battery system is performed using electrical, thermal, and aging sub-models. Dynamic updates of model parameters are achieved by combining a dual-frequency strategy and extended Kalman filtering. Adaptive gating parameters are introduced to dynamically fuse the outputs of the physical model and the data-driven model. High-precision state estimation is achieved through dynamic weighted fusion and residual compensation mechanisms. A multi-level structure mapping engine is constructed to realize bidirectional mapping between different systems such as cells, packs, clusters, and modules. A multi-scale digital twin model is also constructed.

[0125] The trend simulation and state prediction module, in which:

[0126] Based on a multi-scale digital twin model, a coupled state mechanism of electro-thermal-aging is constructed. Through cross-scale state convergence and feature reconstruction, multi-level state co-evolution is achieved. A predictive baseline for state evolution is generated through a physical model, and a data-driven model is used to correct physical prediction biases. A multi-time-domain anomaly judgment system is constructed, and the physical and data residuals of key monitoring indicators are calculated. Combined with rate of change detection, time-series anomaly detection models, and long-term trend analysis, anomaly localization is achieved through an electro-thermal-aging collaborative feedback mechanism. Through feature contribution quantification, counterfactual simulation, and feature attribution analysis, combined with attention mechanisms and causal analysis, multi-level explanatory information is generated.

[0127] The visualization and intelligent early warning module includes:

[0128] A three-dimensional spatial topology model of cell-pack-cluster-module is established. The cell status is aggregated to the PACK level, cluster level, and module level through hierarchical mapping rules, maintaining the topological relationship in physical space. Key status parameters are dynamically bound to visual primitives, and the values ​​are mapped to visual attributes through normalization, weighting, and topological constraints. Key anomaly indicators are recorded and visualized in time series. Alarm signals are generated based on trend prediction results and anomaly trajectories and are processed in a hierarchical manner, using adaptive threshold calculation, multi-dimensional risk scoring, and historical pattern comparison.

[0129] The beneficial effects of this invention lie in its construction of a high-precision, interpretable, predictive, and visualized digital twin system for energy storage batteries. Addressing the problems of inaccurate state modeling, insufficient sensing granularity, uninterpretable predictions, and lag in existing energy storage systems, this invention establishes a multi-level, multi-dimensional virtual mapping system by integrating physical mechanism models and data-driven methods. This enables real-time sensing, dynamic simulation, and trend prediction of battery operating status. A cloud-edge-device architecture is adopted for the state simulation and prediction of the power station system. At the edge, the instantaneous state, corrected by physical-data fusion, is obtained, ensuring data accuracy and continuity. At the cloud, interpretable trend prediction and anomaly identification are achieved, and adaptive alarm signals are generated based on the prediction results. Combined with multi-dimensional risk scoring, tiered early warnings and operation and maintenance suggestions are provided.

[0130] The system employs modeling approaches at different granularities, such as cells, packs, clusters, and compartments, achieving multi-level mapping from cells to packs to compartments. It combines multi-physics coupling models (thermal, electrical, and aging) with machine learning algorithms to improve the accuracy and verifiability of predicting battery performance evolution. Simultaneously, it integrates a 3D visualization module, using a state-structure binding mechanism to map multi-level simulation data to compartment / cluster / pack / cell topology models. Through dynamic colors, deformations, and animations, it intuitively displays battery status, operating trends, and abnormal evolution processes, enhancing system interactivity and interpretability. Based on this, the system can autonomously perform functions such as status monitoring, anomaly warnings, and predictive maintenance recommendations, providing integrated intelligent simulation and proactive management capabilities for energy storage systems.

[0131] Therefore, it is evident that the present invention has outstanding substantive features and significant progress compared with the prior art, and the beneficial effects of its implementation are also obvious. Attached Figure Description

[0132] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0133] Figure 1 This is a schematic flowchart of a method for simulating and predicting the state of energy storage batteries based on digital twins according to the present invention.

[0134] Figure 2 This is a system architecture diagram of a system for simulating and predicting the state of energy storage batteries based on digital twins, according to the present invention. Detailed Implementation

[0135] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. The following embodiments are explanations of the present invention, but the present invention is not limited to the following implementation methods.

[0136] Example 1:

[0137] This invention provides a method for simulating and predicting the state of energy storage batteries based on digital twins, specifically including the following steps:

[0138] Step S1, the step of constructing a multi-scale digital twin model, in which:

[0139] Multiphysics modeling of the battery system is performed using electrical, thermal, and aging sub-models; a dual-frequency strategy and extended Kalman filtering are employed to achieve dynamic updating of model parameters; adaptive gating parameters are introduced to dynamically weight and fuse the outputs of the physical model and the data-driven model, combined with a residual compensation mechanism to achieve high-precision state estimation; a multi-level structure mapping engine is constructed to achieve bidirectional mapping between cells, packs, clusters, and modules; and a multi-scale digital twin model is built.

[0140] Step S2, the trend simulation and state prediction step, in which:

[0141] Based on the multi-scale digital twin model constructed in step S1, a coupled state mechanism of electro-thermal-aging is built. Through cross-scale state convergence and feature reconstruction, multi-level state co-evolution is achieved. A predictive baseline for state evolution is generated through a physical model. The data-driven model is used to correct physical prediction biases. A multi-time-domain anomaly judgment system is constructed. The physical residuals and data residuals of key monitoring indicators are calculated. Combined with rate of change detection, time-series anomaly detection models, and long-term trend analysis, anomaly localization is achieved through an electro-thermal-aging collaborative feedback mechanism. Through feature contribution quantification, counterfactual simulation, and feature attribution analysis, combined with attention mechanisms and causal analysis, multi-level explanatory information is generated.

[0142] Step S3, the visualization and intelligent early warning step, in which:

[0143] A three-dimensional topology model of cell-pack-cluster-module is constructed, and the status data is bound to the virtual model; key status parameters are dynamically mapped through color gradients, geometric deformations, etc.; the time series of key abnormal indicators are recorded, and dynamic playback and trend analysis are performed in combination with the three-dimensional topology model; hierarchical alarm signals are generated based on prediction results and abnormal trajectories, and operation suggestions are provided in combination with historical patterns and expert rules, and the affected areas are highlighted in the three-dimensional interface.

[0144] The steps for constructing the multi-scale digital twin model further include the following steps:

[0145] S101, Steps for physical mechanism modeling, in which:

[0146] Equivalent thermal model, electrochemical model and aging mechanism model are used to simulate the response behavior of battery under different dimensions. Through electrical sub-model, thermal sub-model and aging sub-model, the current-voltage dynamic response, heat conduction and heat dissipation process and capacity decay and internal resistance growth law of battery are modeled respectively.

[0147] Furthermore, the electrical sub-model adopts a second-order Thevenin equivalent model to dynamically model the cell terminal voltage and polarization voltage.

[0148] Furthermore, the inputs to the electrical sub-model include real-time current, voltage, ambient temperature, initial SOC, and initial voltage polarization state; model parameters, including ohmic internal resistance, electrochemical polarization resistance, and capacitance, are obtained through experimental calibration or historical data fitting.

[0149] Furthermore, during the operation of the electrical sub-model, a dual-frequency strategy is adopted to calculate and update the parameters to obtain output parameters, including real-time SOC, terminal voltage prediction, and polarization voltage components. The output here not only provides the basis for heating calculation for the subsequent thermal sub-model, but also provides electrical state reference for the aging sub-model and the data-driven model, supporting multi-physical quantity coupling analysis and cell-level performance prediction.

[0150] Furthermore, in the dual-frequency strategy, the high frequency indicates that the model parameters remain unchanged in the short term. The current parameters are used to simulate the polarization effect inside the cell in real time through a second-order RC model to calculate parameters such as voltage polarization component, terminal voltage, and SOC. The low frequency indicates that under fixed period or triggering conditions, the historical parameters and operational observation data are used to correct the model parameters through extended Kalman filtering to obtain updated impedance, capacitance, and polarization state parameters.

[0151] Furthermore, the thermal sub-model adopts a hierarchical thermal network structure, in which the cell level adopts a multi-node thermal model, and the PACK, cluster, and compartment system levels adopt equivalent thermal node models to simulate the thermal coupling and temperature distribution between cells, PACK, clusters and compartments.

[0152] Furthermore, the inputs to the thermal sub-model are the output voltage, SOC, real-time current, ambient temperature, and cooling medium state of the electrical sub-model; model parameters, including cell specific heat capacity, thermal conductivity, and convective heat transfer coefficient, are obtained through experimental measurement or calibration.

[0153] Furthermore, in the process of processing, each cell is abstracted into several hot nodes, each node representing a physical region of the cell; the node parameters are assigned according to the physical location, and the temperature of each node of the cell is calculated using the energy conservation and heat conduction equations. Taking into account the Joule heat generated by the current, the heat conduction between adjacent cells and external heat dissipation, the hot spot locations of the temperature distribution between the cell and the compartment are obtained.

[0154] Furthermore, the Joule heating formula is as follows:

[0155]

[0156] Where Q is Joule heat, I is current, and R is resistance;

[0157] Furthermore, the node parameters include specific heat, thermal conductivity, and convective heat transfer area, and the physical region includes key heat capacity units such as positive electrode plates, negative electrode plates, electrolyte, or shell.

[0158] Furthermore, the output of the thermal sub-model is cell temperature and hotspot information, which provides temperature input for the aging sub-model and can be used for twin visualization mapping.

[0159] Furthermore, at the PACK, cluster, and cabin levels, an equivalent single-node thermal model of the battery cell is used for modeling. The next-level structure is thermally equivalent to a lumped parameter node, used to describe the overall average temperature state and heat exchange behavior of that level. The heat capacity of the equivalent single node is obtained by weighting the heat capacities of its subordinate units by quantity or mass, and the equivalent thermal resistance is determined by the structural connection relationship, cooling path, and installation method.

[0160] Furthermore, the aging sub-model models cell capacity decay and SOH evolution based on empirical or semi-physical equations.

[0161] Furthermore, the inputs to the aging sub-model include cell SOC, temperature, charge-discharge cycle count, and historical capacity decay data. The model parameters, including the aging rate constant and decay coefficient, are obtained through experimental fitting.

[0162] Furthermore, the aging sub-model calculates capacity loss based on conditions such as SOC, temperature, and cycle count, and predicts the remaining capacity and SOH evolution trend of the cell. The specific expression is as follows:

[0163]

[0164] in, This indicates the degree of degradation of the battery's health status per unit time. The larger the value, the faster the battery ages. This indicates the battery's state of charge, a key parameter affecting aging. Representing temperature and charge / discharge cycle count, respectively, the function... It can be implemented by including a temperature Arrhenius term and a load / cycle term, so that the aging rate increases with increasing temperature, cumulative cycles, and enhanced high load conditions;

[0165] Furthermore, the output of the aging sub-model includes the cell's remaining capacity, SOH, decay rate, and capacity inconsistency index, providing a physical basis for trend prediction, anomaly identification, and multi-scale twin mapping.

[0166] Furthermore, the capacity inconsistency index is calculated from the SOH distribution of the cell group, including range, standard deviation or quantile difference;

[0167] Furthermore, to improve the online accuracy and adaptability of the physical model, the physical mechanism introduces an extended Kalman filter (EKF) to update the cell state at high and low frequencies. The inputs of the EKF include the predicted states output by the electrical, thermal, and aging sub-models, as well as real-time measurements. The state estimation increment is calculated through the state equation and the measurement equation.

[0168] Furthermore, the EKF employs a dual-timescale update strategy: high-frequency updates are used to perform real-time corrections on fast states such as SOC, polarization components, and temperature to quickly track operational fluctuations; low-frequency updates are used to recursively estimate slowly changing parameters, including polarization resistance / capacitance, equivalent thermal parameters, and aging rate parameters, enabling the model to self-correct as operating conditions and life stages change.

[0169] Furthermore, the predicted state includes SOC, terminal voltage, polarization voltage, temperature, and capacity loss, and the real-time measured values ​​include current, voltage, and ambient temperature.

[0170] Furthermore, the model parameters include polarization resistance, capacitance, thermal conductivity, and aging rate;

[0171] Furthermore, the EKF outputs the corrected cell-level SOC, voltage polarization component, temperature, and SOH, and generates derived features such as capacity imbalance and hot spots, providing a reliable physical benchmark for subsequent data-driven compensation, anomaly prediction, and visualization.

[0172] Step S102, the physical-data fusion modeling step, in which:

[0173] By introducing adaptive gating parameters constrained by physical consistency into the model structure, and integrating the mechanistic and data characteristics of cell operation, the problem of insufficient modeling accuracy and difficulty in characterizing nonlinear features under complex operating conditions of the physical model is solved. This balances physical consistency and nonlinear expressive ability, thereby achieving accurate estimation of key state quantities such as SOC, SOH, terminal voltage and temperature under complex operating conditions, and providing a stable and reliable state benchmark for subsequent trend prediction, anomaly identification and multi-level twin mapping.

[0174] Furthermore, the inputs to the steps include cell SOC, terminal voltage, polarization voltage, temperature, and SOH obtained from physical mechanism modeling, as well as historical operating data, including charge and discharge current, voltage, temperature, depth of charge and discharge (DOD), rate, number of cycles, and operating condition markers, etc.

[0175] Furthermore, a baseline state equation is constructed based on the electrical sub-model, thermal sub-model, and aging sub-model to characterize the dynamic evolution of the cell under the influence of current, temperature, and aging. The baseline state equation serves as the physical constraint and prediction reference for the fusion model, and is used to limit the output space of the data-driven model.

[0176] Furthermore, after the running data is interpolated, normalized, and anomaly detected, multidimensional time series features are formed based on the time window division, and derived indicators such as voltage change rate, temperature gradient, and capacity decay rate are constructed to enhance the model's ability to describe nonlinear features.

[0177] Furthermore, in the data-driven stage, a Long Short-Term Memory (LSTM) network is used to capture the dynamic dependence and cross-time characteristic coupling of the cell during the cycling process. Through a gating mechanism, the LSTM is used to characterize complex nonlinear relationships such as SOC-voltage and temperature-degradation.

[0178] Furthermore, during the training process, long-term dependency information is stored and updated through a time-step expansion gating mechanism. The loss function is optimized primarily by the mean square error of the predicted terminal voltage, SOC, and temperature. Capacity inconsistency index and temperature gradient features are introduced as auxiliary constraints to improve the fitting ability to boundary conditions and atypical modes.

[0179] Furthermore, after the model training is completed, a short-term prediction of the future state of the battery cell is generated based on the current state sequence.

[0180] Furthermore, in the model fusion stage, a gating function g(t) is introduced to control the physical model prediction and data-driven prediction. Dynamic weighted fusion is achieved by adaptively adjusting model weights. The gate function takes the current SOC, temperature, aging degree, and predicted residual characteristics of the battery cell as input. It dynamically calculates the weight allocation relationship between the physical model and the data-driven model through nonlinear mapping, thereby realizing adaptive switching and collaborative output of the two types of models at different operating stages. The fusion form is as follows:

[0181]

[0182] Among them, g(t) is a gate parameter, which is obtained by nonlinear mapping calculation based on the real-time state of the battery cell. It can dynamically adjust the contribution ratio of the two types of models according to different operating conditions. The predicted values ​​are from the data-driven model. These are the predicted values ​​from the physical model; This is the final fusion prediction value;

[0183] Furthermore, when the system is in a steady state or linear range, the physical model dominates; while in the stages of strong nonlinearity, abrupt feature changes, or aging, the data-driven model is used for correction and compensation, thereby achieving a dual guarantee of prediction accuracy and physical consistency.

[0184] Furthermore, to reduce model bias and improve long-term operational stability, the system introduces a residual compensation mechanism: by calculating the difference between the physical prediction and the fused prediction and applying a dynamic correction coefficient α(t), the fusion result is corrected a second time, as shown in the following expression:

[0185]

[0186] in, This is a dynamic correction factor; It is the final accurate prediction result output by the system. This is to connect with the intermediate prediction results mentioned earlier.

[0187] Furthermore, the final output cell's core state parameters at a given moment, such as SOC, terminal voltage, polarization voltage, temperature, and SOH, are generated simultaneously, along with derived characteristics such as capacity inconsistency, hotspot indicators, and degradation rate. These outputs serve as standard inputs for the multi-level twin system, providing accurate and interpretable foundational support for multi-level twin mapping, trend prediction, and anomaly identification.

[0188] Step S103, the steps of the multi-level structure mapping engine, specifically also include:

[0189] A multi-level structure mapping engine for energy storage systems is constructed to realize cross-scale state mapping and feature reconstruction of cells-packs-clusters-modules in energy storage systems. The mapping engine takes the cell-level fused state as the basic input and establishes a continuous association from the micro-cell state to the macro-system performance by introducing electrical topology constraints, thermal coupling relationships and physical attribute weights, thereby achieving data consistency, physical constraint consistency and feature traceability.

[0190] Furthermore, the multi-level structure mapping engine adopts a bidirectional mapping mechanism of "uplink aggregation - downlink reconstruction". Uplink aggregation is used to aggregate the cell-level state to the PACK, cluster and module levels step by step. Downlink reconstruction is used to reverse calculate the cell-level state based on historical mapping relationships and weight distribution when system-level analysis or anomaly localization is required, thereby realizing bidirectional state synchronization across scales and providing a unified state expression basis for digital twin systems.

[0191] Furthermore, the inputs to the multi-level structure mapping engine include key state variables output from the cell-level fusion model, as well as derived features generated by the data-driven module, while also incorporating environmental layer information to enhance the integrity and physical accuracy of cross-level mapping.

[0192] Furthermore, the key state quantities include SOC, terminal voltage, polarization voltage, temperature, and SOH; the derived characteristics include capacity inconsistency, temperature difference hotspots, aging rate, etc.; and the environmental layer information includes liquid cooler unit status, ambient temperature, air-cooled / liquid-cooled medium parameters, etc.

[0193] Furthermore, the input data is standardized, dimensionally normalized, and weighted. An adaptive weighting mechanism based on physical attributes and spatial location is introduced. The weighting coefficients can be adaptively adjusted according to the importance of the cells, their location distribution, and their historical health status, so that the characteristics of different cells have a unified dimension and numerical range, ensuring numerical comparability and computational stability during multi-level aggregation, so as to reflect the overall performance.

[0194] Furthermore, during the uplink aggregation stage, the multi-level structure mapping engine aggregates cell-level data into PACK-level, cluster-level, and compartment-level states based on the electrical and thermal topology of the battery pack, including series and parallel relationships, physical spatial location, and thermal coupling paths.

[0195] The convergence process employs a multi-parameter weighting strategy, during which:

[0196] Voltage aggregation follows the series and parallel circuit rules, combining layer by layer through the topology matrix to obtain the total voltage of the PACK, cluster, and compartment. SOC aggregation adopts a capacity-weighted average method, with the PACK-level SOC calculated using the effective capacity of individual cells as the weight, and the cluster and compartment-level SOC iteratively aggregated using the rated capacity of the PACK and cluster levels as the weight. Temperature aggregation is based on a heat capacity-weighted average, and simultaneously calculates consistency indicators such as maximum temperature difference and standard deviation to reflect the degree of thermal imbalance within the PACK, cluster, and compartment. When a State of Health (SOH) aggregation is performed, a capacity-weighted strategy or a conservative worst-case strategy can be adopted according to application requirements to achieve different orientations for performance characterization or safety characterization.

[0197] Furthermore, to enhance the model's engineering robustness, the mapping engine simultaneously calculates consistency metrics for each level during the convergence phase, including voltage standard deviation, SOC dispersion, and temperature range. Based on these consistency metrics, it establishes inconsistency criteria between levels to detect potential individual cell failures or system anomalies. If missing or abnormal cell data is detected, the engine completes the state through a multi-source feature reconstruction algorithm.

[0198] Furthermore, the completion process can comprehensively adopt the following methods: interpolation correction based on the thermal coupling relationship of adjacent cells; dynamic compensation based on time series models; and weighted estimation based on the historical distribution pattern of cells, thereby achieving data continuity and spatial integrity without increasing the cost of additional sensors.

[0199] Furthermore, during the downlink reconstruction phase, when the system needs to perform detailed analysis or local state deduction at the PACK or cell level, the mapping engine can reverse calculate the cell-level state based on the historical mapping matrix and weight distribution, supporting single-unit hotspot backtracking, health state reconstruction and local anomaly tracking, and realizing bidirectional state mapping across the entire link.

[0200] Furthermore, the final output of the system includes the cell-level status after downlink reconstruction, and the PACK-level, cluster-level, and compartment-level status after uplink aggregation, specifically covering the SOC, voltage, temperature, aging status, and derived abnormal indicators at each level; providing standardized input for trend prediction, anomaly identification, and 3D visualization modules;

[0201] Step S1 addresses the technical problem that traditional modeling methods cannot simultaneously consider the nonlinearity, multi-physics coupling characteristics, and fine-grained state mapping of battery systems, making it difficult to construct a high-precision, real-time responsive virtual mapping entity. The following solutions are provided: A hybrid modeling method based on multi-source fusion is proposed. This method introduces a data-driven model under physical model constraints to compensate for nonlinear relationships such as SOC-voltage and temperature-aging. Adaptive gating parameters are used to dynamically adjust the weights of the two types of models, effectively compensating for modeling errors in pure physical models under conditions of strong nonlinearity and significant aging. This approach balances physical constraints and data learning, improving simulation accuracy and dynamic adaptability. A unified virtual mapping of multi-granularity and multi-dimensional states is achieved, supporting fine-grained state reproduction from cell to system level. This ensures physical consistency and data continuity in state evolution across different levels, providing a reusable and scalable foundational model for multi-level digital twin mapping. Online model calibration and adaptive lifecycle evolution update mechanisms are supported. Through residual compensation and dynamic parameter update mechanisms, the model can continuously correct prediction deviations based on real-time data during operation, possessing long-term evolution capabilities.

[0202] Step S2 specifically also includes:

[0203] Step S201: A step in the multi-field coupled prediction algorithm, in which:

[0204] The core idea is to leverage the complex coupling relationship between physical, thermal, and electrochemical aging for multi-time-domain prediction. A multi-physics coupling prediction algorithm is employed to simulate and predict the state evolution of energy storage batteries across multiple time domains. The algorithm uses the cell-level fusion state output in step S1 as initial conditions and simultaneously describes the coupling feedback relationship between electrical dynamics, thermal transfer, and aging evolution within a unified state space. This addresses the problems of unstable predictions and unreliable long-term trends under complex operating conditions caused by single models. The algorithm's inputs include cell-level state vectors, historical observation sequences, and future operating condition assumptions (load plans, temperature control strategies, ambient temperature boundaries, etc.). The outputs include predicted trajectories, confidence intervals, and risk metrics across multiple levels from cell to PACK to cluster to module.

[0205] Furthermore, the dynamic evolution of the cell terminal voltage and polarization component as a function of current fluctuation is calculated using an electrical sub-model. At the same time, the thermal sub-model is combined to simulate the heat conduction, convection heat transfer and external heat dissipation effects of the cell, PACK, cluster, and inter-cell to obtain the temperature distribution.

[0206] Furthermore, the aging sub-model updates SOH and related aging indices based on slow variables such as temperature and cyclic load dynamics, and uses the aging state feedback to correct the electrical and thermal model parameters, forming a closed-loop coupling mechanism to achieve real-time interaction and dynamic adjustment between different physical quantities. It maps the change in SOH / internal resistance to the drift of ohmic internal resistance and polarization impedance, and introduces the change in heat exchange efficiency or thermal resistance into the update of thermal model parameters, thereby forming a closed-loop coupling link of "electric drive - thermal response - aging accumulation - parameter drift - electrothermal re-response", so that the mutual influence between different physical quantities can be continuously reflected in the prediction process.

[0207] Furthermore, in implementation, the battery cell is the smallest computational unit, and the state vector of each cell includes SOC, terminal voltage, polarization components, temperature, and SOH. A uniform time step is used for prediction. The state equations are iteratively updated, forming a multi-step forward propagation process: In each time step, the terminal voltage and polarization components are first updated by the electrical model based on the given operating conditions (current, cooling state, environmental boundary), then the temperature state is updated by the thermal model based on the heat generation power and thermal network, and finally the SOH / internal resistance drift is updated by the aging model based on the temperature and cumulative load, and this drift is written back to the electrical and thermal parameters of the next step, realizing the dynamic self-consistent update of the coupling parameters. To ensure cross-scale consistency, the PACK, cluster, and compartment-level states are obtained in real time from the cell-level states through topological constraints and weighted rules, and are used to output system-level prediction results and consistency indicators, ensuring cross-scale data consistency.

[0208] Furthermore, the generation of the physical prior trajectory is achieved through multi-step forward propagation; the current cell state is used as the initial condition, and the discretized electro-thermal-aging coupling equation is iteratively calculated at a fixed time step to obtain the short-term prediction results, forming a physical baseline;

[0209] Furthermore, in short-term forecasts at the minute to hour level, high-frequency sampling data drives the dynamic updating of the physical model to ensure real-time responses to terminal voltage, SOC, and temperature. In long-term forecasts at the day to week and month to year level, since high-frequency measurements are unavailable, the system extends the physical baseline trajectory through historical cyclic statistics, typical load sequences, and low-frequency parameter corrections (such as temperature-dependent aging rates and internal resistance growth curves), making the long-term forecasts trend-based and providing a basic reference for residual learning.

[0210] Furthermore, historical observation sequences are combined with physical baseline residuals to capture nonlinear behaviors that physical models cannot fully reflect. The residual sequences are encoded in time series networks (such as LSTM, GRU, or lightweight Transformer) and combined with historical state features, including current change rate, SOC gradient, temperature distribution features, and aging accumulation, to learn systematic biases under electro-thermal-aging coupling.

[0211] Furthermore, the loss function during network training not only includes the mean square error of terminal voltage, SOC, and temperature, but also introduces physical consistency constraints (charge conservation, SOH monotonicity) and uncertainty constraints (NLL or quantum loss) to ensure that residual predictions can capture boundary conditions and atypical patterns while conforming to physical laws. In the prediction stage, the output residual sequence is combined with the physical baseline to generate multi-time-domain prediction results, achieving a closed-loop fusion of "physical prior + data residuals".

[0212] Furthermore, the fusion of multi-dimensional states adopts a unified cell-level state vector representation. In each time step, current and voltage update cell polarization and terminal voltage, the thermal model updates the individual cell temperature and calculates heat conduction and dissipation of adjacent cells, and the aging model updates SOH and internal resistance based on temperature and load. The aging state is then fed back to correct electrical and thermal parameters, forming a closed-loop feedback across physical fields. In PACK-level, cluster-level, and compartment-level convergence, SOC uses a capacity-weighted average, temperature uses a thermal capacity-weighted average, and SOH can be either capacity-weighted or a worst-case strategy. Combined with topological constraints and position weights, cross-scale state consistency is ensured. When missing or abnormal cell data exists, a feature reconstruction algorithm is introduced to recover the state, ensuring a continuous and robust prediction process.

[0213] Furthermore, through the above process, continuous state simulation and trend prediction from short-term minute-level to long-term month-level levels are achieved, while outputting confidence intervals and risk metrics to support operation and maintenance decisions and preventive maintenance. The innovation of this method lies in constructing an electro-thermal-aging closed-loop coupling by combining physical priors and residual correction. At the same time, it utilizes multi-time-domain residual learning and cross-scale state convergence to achieve full-process virtual simulation and trend prediction. This ensures that the prediction results maintain physical rationality while possessing nonlinear adaptability and engineering interpretability, providing a reliable foundation for subsequent abnormal state prediction and intelligent early warning.

[0214] Step S202, the steps of the abnormal state prediction mechanism, in which:

[0215] The aim is to identify potential hidden faults and risks in energy storage battery systems in advance, and to address the shortcomings of traditional prediction methods in anomaly detection and hidden fault identification due to their reliance on thresholds or single models.

[0216] Furthermore, based on the fusion prediction trajectory (SOC, terminal voltage, polarization component, temperature, SOH and derived indicators) output in step S1, this module combines historical observation data and future operating condition assumptions to construct a multi-time-domain anomaly judgment system, realizing a multi-time-domain anomaly prediction and graded risk quantification system from individual cells to PACKs and even clusters and modules.

[0217] Furthermore, to fully capture potential abnormal behavior, anomaly information is first modeled. Based on the fused predicted SOC, terminal voltage, polarization components, temperature, SOH, and their derived characteristics, a key monitoring index system is constructed. These indicators include at least: cell terminal voltage and group dispersion, single-cell equivalent internal resistance or voltage-SOC offset rate, temperature difference and temperature rise rate between single cell and PACK / cluster, and aging rate deviation. For each index, physical residuals (observed value minus physical baseline predicted value) and data residuals (observed value minus fused predicted value) are calculated to reveal deviations that the physical model and data model cannot fully explain. The residual sequence forms a multi-scale signal in the time dimension. Anomaly characteristics are identified by analyzing its instantaneous changes, cumulative trends, and pattern evolution to characterize abnormal deviation behaviors that are difficult for both the physical and data models to explain.

[0218] Furthermore, through residual sequence analysis, the module can perform multi-time-domain anomaly identification at different time scales, including short-term, medium-term, and long-term. Short-term anomaly identification focuses on sudden events at the minute to hour level, employing rate of change detection and statistical thresholding methods, including exponentially weighted moving average (EWMA), cumulative sum (CUSUM), and Bayesian online change point detection. These algorithms can calculate the residual change rate in real time, capture rapid abnormal fluctuations in voltage or temperature, and achieve highly sensitive alarms.

[0219] Furthermore, mid-term anomaly identification focuses on the development of inconsistencies at the day-week level, and uses time-series anomaly detection models, such as LSTM-Autoencoder to reconstruct the residual sequence and calculate the reconstruction error, and IsolationForest to score anomalies in multidimensional states, thereby discovering gradually accumulating deviations or anomaly patterns.

[0220] Furthermore, long-term anomaly identification targets potential failure risks from month to year. Based on RUL estimation, Bayesian survival analysis, and trend regression, it integrates cell aging status, temperature history, and capacity decay rate to predict possible failure time windows and changes in healthy life.

[0221] Furthermore, in terms of multi-physics collaborative analysis, the anomaly prediction mechanism maintains consistency with the electro-thermal-aging closed-loop coupling model in step S201. Polarization changes caused by current and voltage disturbances, the amplification effect of temperature distribution on local aging rates, and the reverse correction of electrical and thermal parameters by aging states are all incorporated into the anomaly determination process, enabling anomaly identification not only based on apparent data deviations but also reflecting the inherent imbalance mechanism under multi-physics coupling.

[0222] Furthermore, multi-dimensional information such as voltage, current, temperature, and SOH are mapped into a unified state vector, which is dynamically updated iteratively over time steps. Combined with topological constraints, anomaly detection results are located at the specific cell, PACK, cluster, and compartment levels, achieving graded risk quantification. The anomaly detection threshold and model parameters adopt an adaptive update strategy, automatically adjusting based on historical normal segment statistics or rolling quantiles, enabling the system to adapt to changes in operating conditions and data drift, while reducing the false alarm rate.

[0223] Furthermore, the module output includes abnormal indicator values, risk levels, potentially affected individual units, packs or clusters, and estimated remaining response time, providing quantitative basis for graded alarms and preventive maintenance.

[0224] Furthermore, through multi-time-domain and multi-level anomaly identification and prediction, the system can provide actionable risk information before anomalies occur, enabling the operation and maintenance team to respond quickly to local anomalies in the short term, and formulate maintenance strategies or adjust load / temperature control schemes in the medium and long term planning to ensure the safe and stable operation of the energy storage system.

[0225] Step S203, the step of the interpretability enhancement mechanism, in which:

[0226] The interpretability enhancement mechanism aims to transform the output of the fusion prediction model from a black-box result into traceable, quantifiable, and operationally guiding explanatory information, addressing the problem that existing data-driven predictions struggle to explain the basis for predictions and cannot support risk assessment and intervention decisions. This mechanism constructs an interpretable closed loop throughout the entire prediction process, from input features to prediction results, and then to risk sources and intervention strategies.

[0227] Furthermore, this mechanism constructs a complete interpretable prediction closed loop through input processing, feature contribution quantification, counterfactual simulation, feature attribution, and fusion display. Its inputs include the instantaneous state of the cell (SOC, SOH, terminal voltage, polarization voltage, temperature distribution, etc.) output by the fusion model in step S1, historical observation sequences, residual sequences, and future operating condition assumptions (load curve, temperature control strategy, charge and discharge rate, etc.), providing basic data for subsequent interpretable analysis.

[0228] Furthermore, in the input processing stage, the mechanism first normalizes, temporally windows, and performs hierarchical mapping on the input features to maintain the correspondence between each input variable in the time and spatial dimensions (cell / pack / cluster / module).

[0229] Furthermore, by utilizing the attention or gating mechanism within the time-series prediction model, the contribution weight of the input features at each time step to the final prediction result is calculated. Specifically, in LSTM, GRU, or lightweight Transformer networks, gradient backpropagation is performed on the hidden state and output, combined with the attention weight matrix, to quantify the impact of each input feature on the prediction target at different time steps into a numerical index. This contribution can be used to generate time-feature heatmaps or directly to evaluate key driving factors.

[0230] Furthermore, after the feature contribution quantification is completed, counterfactual simulation is introduced to evaluate the causal impact of a single factor on the predicted output. This is achieved using a digital twin model, where one or more input variables (such as future temperature, peak load, and discharge rate) are adjusted by a set amplitude, while keeping other conditions constant. The fused prediction model is then rerun to generate a new prediction trajectory. By comparing the difference between the original prediction and the counterfactual trajectory, the specific impact of the factor on SOC, SOH, temperature evolution, or RUL can be quantified, thus achieving causal interpretability.

[0231] Furthermore, feature attribution analysis is performed on the final fusion prediction results. Using approximation schemes such as SHAP or Integrated Gradients, the fusion prediction output is decomposed into the sum of the contributions of each input feature, including both the physical baseline and the residual correction. By performing feature attribution on the residual sequence, the relative contributions of factors such as current fluctuations, temperature changes, historical residuals, or aging conditions to prediction bias and risk scores can be clearly identified, providing quantitative evidence for operations and maintenance personnel.

[0232] Furthermore, in the output phase, the mechanism integrates attention / feature contribution, counterfactual simulation results, and attribution analysis to generate multi-level explanatory information, including time-feature level contribution heatmaps, causal sensitivity analysis, residual source decomposition, confidence intervals, and risk levels. By combining this with topological information, it can locate cells, PACK units, clusters, or modules where anomalies may occur and provide actionable intervention suggestions. Ultimately, this mechanism can provide complete, traceable, quantifiable, and cross-scale predictive explanations, explaining not only "what might happen in the future" but also clarifying "why it might happen" and "how to intervene," significantly improving the credibility and engineering deployability of the prediction results.

[0233] Step S2 addresses the challenges of complex state changes in energy storage battery systems. Traditional methods struggle to accurately predict future states (insufficient prediction accuracy under complex operating conditions), lack interpretability in data-driven prediction processes, and fail to provide reliable data for operation and maintenance decisions. The following solutions are proposed: Implement full-process virtual simulation and trend prediction of operating states and health parameters to overcome the shortcomings of traditional post-event response; introduce a multi-physics collaborative prediction mechanism to enhance the predictive model's adaptability to degradation behavior under complex operating conditions; and adopt an interpretable modeling framework. By quantifying the contribution of input features and decomposing residual sources in the fusion prediction model, the role of SOC, temperature, aging state, and historical operating characteristics in the prediction results can be clearly defined, providing clear physical and data-driven evidence for the prediction conclusions and improving the credibility and engineering applicability of the prediction results.

[0234] Step S3 specifically also includes:

[0235] Step S301, the steps of the three-dimensional structure binding mechanism, in which:

[0236] A three-dimensional spatial topology model of cell-pack-cluster-module consistent with the actual energy storage system is established. Each cell, pack, cluster, and module corresponds to a unique entity in the virtual three-dimensional environment, and its spatial coordinates, hierarchical membership, and topological connection relationships are maintained. The module input is the real-time status of the cell level (SOC, SOH, temperature, polarization component, capacity inconsistency, etc.) and derived characteristics output from steps one and two.

[0237] Furthermore, through hierarchical mapping rules, the cell status is not only aggregated at the PACK, cluster, and module levels, but also maintains its topological relationship in physical space, enabling cross-scale synchronous state tracking. Each visualized object is fully mapped to the actual entity in terms of three-dimensional coordinates, hierarchical structure, and associated attributes. The visualized object, as a state-bearing node, includes at least SOC, SOH, temperature, terminal voltage, polarization components, and capacity inconsistency indicators. This solves the problem that traditional visualization remains at the macro level and cannot present the dynamic evolution of the cell, providing maintenance personnel with a fine-grained foundation for state awareness.

[0238] Step S302, the steps of the parameter dynamic mapping algorithm, in which:

[0239] In the parameter dynamic mapping algorithm, key state parameters (temperature, voltage, capacity decay rate, SOH, etc.) are dynamically bound to visual primitives to achieve intuitive feedback such as color gradients, geometric deformations, and animated flashing. Inputs include real-time state data, historical trajectory sequences, and user-defined mapping rules.

[0240] Furthermore, during the processing, numerical values ​​are mapped to visual attributes through normalization, weighting, and topological constraints, such as color gradients representing temperature, geometric deformation representing capacity decay, and flashing animations indicating abnormal fluctuations, thereby achieving synchronous and dynamic presentation of multidimensional data.

[0241] Furthermore, the algorithm not only supports the simultaneous display of multi-dimensional data, but also reflects the inconsistencies and hotspot locations between battery cells, thereby breaking through the limitations of traditional static display and achieving dynamic visualization with a high degree of consistency between state and structure.

[0242] Step S303, the steps of the abnormal trajectory evolution view, in which:

[0243] The system records and visualizes key anomaly indicators (voltage deviation, temperature rise rate, capacity loss rate, etc.) over time. The input is the state evolution data of each cell; during processing, the indicator trajectories are combined with 3D topology to achieve dynamic playback, trend curve overlay, and local highlighting, supporting source tracing analysis and anomaly evolution tracking.

[0244] Furthermore, this step solves the problem that traditional methods are difficult to analyze the development process of anomalies. Users can intuitively observe the evolution path of anomalies in time and space, support source tracing analysis and anomaly evolution tracking, and provide maintenance personnel with a panoramic view of anomalies from battery cells to PACKs, enabling early perception and analysis of potential risks.

[0245] Step S304, the intelligent early warning step, which generates an alarm signal based on the trend prediction results and abnormal trajectories of step S2 and performs hierarchical processing.

[0246] Furthermore, the intelligent early warning input includes predicted cell status, abnormal indicator trajectories, and an expert rule base. During processing, adaptive threshold calculation, multi-dimensional risk scoring, and historical pattern comparison are employed to provide early warnings for potential thermal runaway, cell inconsistencies, or capacity anomalies. The output includes tiered alarms, affected unit identifiers, and operational suggestions, with affected areas highlighted in a 3D interface. This mechanism solves the problem of slow response times in traditional manual intervention, achieving prediction-driven proactive early warning and enhancing operational efficiency and decision reliability through visualization.

[0247] The above methods overcome the limitations of traditional static displays and manual early warnings. Visual displays not only provide real-time status awareness, but also directly support prediction-driven operation and maintenance and risk management.

[0248] Example 2:

[0249] like Figure 2As shown, the present invention provides a system for simulating and predicting the state of energy storage batteries based on digital twins, including a multi-scale digital twin model construction module, a trend simulation and state prediction module, and a visualization and intelligent early warning module.

[0250] The multi-scale digital twin model construction module further includes:

[0251] Multiphysics modeling of the battery system is performed using electrical, thermal, and aging sub-models. Dynamic updates of model parameters are achieved by combining a dual-frequency strategy and extended Kalman filtering. Adaptive gating parameters are introduced to dynamically fuse the outputs of the physical model and the data-driven model. High-precision state estimation is achieved through dynamic weighted fusion and residual compensation mechanisms. A multi-level structure mapping engine is constructed to realize bidirectional mapping between different systems such as cells, packs, clusters, and modules. A multi-scale digital twin model is also constructed.

[0252] The trend simulation and state prediction module further includes:

[0253] Based on a multi-scale digital twin model, a coupled state mechanism of electro-thermal-aging is constructed. Through cross-scale state convergence and feature reconstruction, multi-level state co-evolution is achieved. A predictive baseline for state evolution is generated through a physical model, and a data-driven model is used to correct physical prediction biases. A multi-time-domain anomaly judgment system is constructed, and the physical and data residuals of key monitoring indicators are calculated. Combined with rate of change detection, time-series anomaly detection models, and long-term trend analysis, anomaly localization is achieved through an electro-thermal-aging collaborative feedback mechanism. Through feature contribution quantification, counterfactual simulation, and feature attribution analysis, combined with attention mechanisms and causal analysis, multi-level explanatory information is generated.

[0254] The visualization and intelligent early warning module specifically includes:

[0255] A three-dimensional topology model of cell-pack-cluster-module is constructed, and the status data is bound to the virtual model; key status parameters are dynamically mapped through color gradients, geometric deformations, etc.; the time series of key abnormal indicators are recorded, and dynamic playback and trend analysis are performed in combination with the three-dimensional topology model; hierarchical alarm signals are generated based on prediction results and abnormal trajectories, and operation suggestions are provided in combination with historical patterns and expert rules, and the affected areas are highlighted in the three-dimensional interface.

[0256] The multi-scale digital twin model construction module is the same as the multi-scale digital twin model construction step S1 in Embodiment 1.

[0257] The trend simulation and state prediction module is the same as the trend simulation and state prediction step S2 in Embodiment 1.

[0258] The visualization and intelligent early warning module is the same as the visualization and intelligent early warning step in step S3 of embodiment 1.

[0259] like Figure 2 As shown, the system adopts a cloud-edge-device architecture for state simulation and prediction of the power plant system. The specific implementation method is as follows:

[0260] The multi-scale digital twin model construction module is deployed on the edge side to collect and preprocess current, voltage, temperature and initial state parameters in real time. Based on the extended Kalman filter, high-frequency state estimation and low-frequency parameter adaptive update are performed on the cell SOC, SOH, polarization voltage and temperature to generate cell-level fine state information. The nonlinear features captured by the data-driven model are supplemented by the gated fusion strategy to finally obtain the instantaneous state after physical-data fusion correction.

[0261] The trend simulation and state prediction module is deployed on the cloud side to receive instantaneous state data from the edge side. It uses thermal-electric-aging joint modeling and time series prediction network to predict the short-term, medium-term and long-term battery operation trends, degradation trajectories and risk indicators. It combines confidence calculation and feature contribution analysis, and generates adaptive alarm signals based on the prediction results. It also provides graded early warning and operation and maintenance suggestions by combining multi-dimensional risk scoring.

[0262] The visualization and intelligent early warning module is integrated into the system. Through the state-structure binding mechanism, multi-level simulation data is mapped to the topology model of the compartment / cluster / PACK / cell. The battery status, operating trend and abnormal evolution process are displayed intuitively through dynamic colors, deformation and animation.

[0263] The above methods overcome the limitations of traditional static displays and manual early warnings. Visual displays not only provide real-time status awareness, but also directly support prediction-driven operation and maintenance and risk management.

[0264] The above-disclosed embodiments are merely preferred embodiments of the present invention, but the present invention is not limited thereto. Any non-creative variations that can be conceived by those skilled in the art, as well as any improvements and modifications made without departing from the principles of the present invention, should fall within the protection scope of the present invention.

Claims

1. A method for simulating and predicting the state of energy storage batteries based on digital twins, characterized in that, Includes the following steps: Step S1, the step of constructing a multi-scale digital twin model, in which: The current-voltage dynamic response, heat conduction and heat dissipation processes, capacity decay, and internal resistance growth of the battery are modeled using electrical, thermal, and aging sub-models, respectively. A dual-frequency strategy and extended Kalman filtering are employed to achieve dynamic updates of model parameters. Adaptive gating parameters are introduced to dynamically weight and fuse the outputs of the physical model and the data-driven model, and state estimation is performed in conjunction with a residual compensation mechanism. A multi-level structure mapping engine is constructed to achieve bidirectional mapping between cells, packs, clusters, and modules. The multi-level structure mapping engine takes the cell-level fusion state as the basic input and establishes a continuous correlation from the micro-cell state to the macro-system performance by introducing electrical topology constraints, thermal coupling relationships and physical attribute weights. The multi-level structure mapping engine adopts a bidirectional mapping mechanism of uplink aggregation and downlink reconstruction. Uplink aggregation is used to aggregate the cell-level state to the PACK level, cluster level and module level step by step. Downlink reconstruction is used to reverse calculate the cell-level state based on historical mapping relationship and weight distribution in the case of system-level analysis or anomaly localization. Step S2, the trend simulation and state prediction step, in which: A coupled state mechanism of electro-thermal-aging is constructed, and multi-level state co-evolution is achieved through cross-scale state convergence and feature reconstruction. A predictive baseline for state evolution is generated through a physical model, and the physical residuals and data residuals of key monitoring indicators are calculated. Anomaly localization is achieved through an electro-thermal-aging co-feedback mechanism. By quantifying feature contributions, counterfactual simulation, and feature attribution analysis, combined with attention mechanisms and causal analysis, multi-level explanatory information is generated. Step S3, the visualization and intelligent early warning step, in which: A three-dimensional spatial topology model of cell-PACK-cluster-module is established, and the cell status is aggregated to the PACK level, cluster level and module level through hierarchical mapping rules; key status parameters are dynamically bound to visual primitives, and the values ​​are mapped to visual attributes through normalization, weighting and topological constraints. Key anomaly indicators are recorded and visualized in time series; alarm signals are generated and graded based on trend prediction results and anomaly trajectories.

2. The method according to claim 1, characterized in that, The electrical sub-model described in step S1 adopts a second-order Thevenin equivalent model. The inputs include real-time current, voltage, ambient temperature, initial SOC, and initial voltage polarization state. The model parameters are obtained through experimental calibration or fitting of historical data. The thermal sub-model adopts a hierarchical thermal network structure, in which the cell level uses a multi-node thermal model, and the PACK, cluster, and compartment system level uses an equivalent thermal node model to simulate the thermal coupling and temperature distribution between the cell, PACK, cluster, and compartment. The inputs are the terminal voltage, SOC, real-time current, ambient temperature, and cooling medium status output by the thermal sub-model. The aging sub-model is based on empirical or semi-physical equations, and its inputs include cell SOC, temperature, charge-discharge cycle count, and historical capacity decay data. The extended Kalman filter employs a dual-time-scale update strategy: high-frequency updates are used for real-time correction of fast states; low-frequency updates are used for recursive estimation of slowly varying parameters.

3. The method according to claim 2, characterized in that, The dynamic weighted fusion described in step S1 introduces a gating function g(t) to adjust the model weights for physical model prediction and data-driven prediction; The gate function takes the cell's current SOC, temperature, aging level, and predicted residual characteristics as input, and dynamically calculates the weight allocation relationship between the physical model and the data-driven model through nonlinear mapping. Its fusion form is as follows: Wherein, g(t) is the gate parameter, which is obtained by nonlinear mapping calculation from the real-time state of the battery cell; The predicted values ​​are from the data-driven model. These are the predicted values ​​from the physical model; This is the final fusion prediction value.

4. The method according to claim 3, characterized in that, In step S1, a residual compensation mechanism is introduced: the fusion result is corrected twice by a dynamic correction coefficient α(t), and the correction formula is as follows: in, This is a dynamic correction factor; It is the final accurate prediction result output by the system. This is to connect with the intermediate prediction results mentioned earlier.

5. The method according to claim 4, characterized in that, In step S1, during the uplink aggregation stage, the multi-level structure mapping engine aggregates cell-level data into PACK-level, cluster-level, and compartment-level states based on the electrical and thermal topology of the battery pack. The aggregation process employs a multi-parameter weighting strategy. Voltage aggregation follows the series and parallel rules of circuits, and obtains the total voltage of PACK, cluster, and compartment by combining them layer by layer through the topology matrix; the multi-level structure mapping engine calculates the consistency index of each level at the same time during the aggregation stage; and establishes the inconsistency criteria between levels based on the consistency index to detect potential failures of individual cells or system anomalies; if some cell data is missing or abnormal, the engine completes the state through a multi-source feature reconstruction algorithm. During the downlink reconstruction phase, when performing detailed analysis or local state simulation at the PACK or cell level, the mapping engine uses the historical mapping matrix and weight distribution to reverse calculate the cell-level state, perform individual-level hotspot backtracking, health status reconstruction, and local anomaly tracking.

6. The method according to claim 5, characterized in that, Step S2 further includes: The dynamic evolution of cell terminal voltage and polarization component with current fluctuation is calculated using an electrical sub-model. At the same time, the thermal sub-model is combined to simulate the heat conduction, convection heat transfer and external heat dissipation effects of the cell, PACK, cluster, and inter-cell to obtain the temperature distribution. Using the cell condition as an initial condition, the discretized electro-thermal-aging coupling equation is iteratively calculated at a fixed time step to obtain short-term prediction results, forming a physical baseline. In short-term forecasting, high-frequency sampling data drives the dynamic updating of the physical model; in long-term forecasting, the physical baseline trajectory is extended by correcting historical cyclic statistics, typical load sequences, and low-frequency parameters. By combining historical observation sequences with physical baseline residuals, nonlinear behaviors that cannot be fully reflected by physical models are captured; the residual sequences are encoded in a time series network, and combined with historical state features, systematic biases under the electro-thermal-aging coupling are learned. In the prediction phase, the output residual sequence is combined with the physical baseline to generate multi-temporal prediction results; A unified cell-level state vector is used to represent the fusion of multi-dimensional states. In each time step, the current and voltage update the cell polarization and terminal voltage, the thermal model updates the individual cell temperature and calculates the heat conduction and heat dissipation of adjacent cells, and the aging model updates the SOH and internal resistance according to the temperature and load, and feeds back the aging state to correct the electrical and thermal parameters. In PACK-level, cluster-level, and compartment-level aggregation, SOC adopts capacity-weighted averaging, temperature adopts thermal capacity-weighted averaging, and SOH adopts capacity-weighted averaging or worst-case strategy, combined with topological constraints and location weights; when there is missing or abnormal cell data, a feature reconstruction algorithm is introduced to restore the state.

7. The method according to claim 6, characterized in that, Step S2 further includes: By employing the attention or gating mechanism within the time-series prediction model, the contribution weight of the input features at each time step to the final prediction result is calculated; the specific implementation method is as follows: In LSTM, GRU, or lightweight Transformer networks, the influence of each input feature on the prediction target at different time steps is quantified into a numerical index by backpropagating the gradients of the hidden state and output, combined with the attention weight matrix. Adjust one or more input variables by a set range, rerun the fusion prediction model to generate a new prediction trajectory, and compare the difference between the original prediction and the counterfactual trajectory. Feature attribution analysis is performed on the final fusion prediction results; using the SHAP and IntegratedGradients approximation schemes, the fusion prediction output is decomposed into the sum of the contributions of each input feature, including the physical baseline part and the residual correction part.

8. The method according to claim 7, characterized in that, Step S3 further includes: Step S301, the steps of the three-dimensional structure binding mechanism, in which: A three-dimensional spatial topology model of cell-PACK-cluster-module consistent with the actual energy storage system is established, and each cell, PACK, cluster, and module corresponds to a unique entity object in the virtual three-dimensional environment. Through hierarchical mapping rules, the cell status is aggregated to the PACK level, cluster level, and module level, while maintaining their topological relationship in physical space. Step S302, the steps of the parameter dynamic mapping algorithm, in which: In the parameter dynamic mapping algorithm, key state parameters are dynamically bound to visual primitives; numerical values ​​are mapped to visual attributes through normalization, weighting, and topological constraints. Step S303, the steps of the abnormal trajectory evolution view, in which: Key anomaly indicators are recorded and visualized over time; the state evolution data of each cell is input, and the indicator trajectory is combined with three-dimensional topology during the processing. Step S304, the intelligent early warning step, in which: Based on the trend prediction results and abnormal trajectories in step S2, alarm signals are generated and processed in a hierarchical manner. The intelligent early warning input includes the predicted cell status, abnormal indicator trajectories, and expert rule base. During the processing, adaptive threshold calculation, multi-dimensional risk scoring, and historical pattern comparison are adopted. The output includes hierarchical alarms, affected unit identifiers, and operation suggestions, and the affected areas are highlighted in the three-dimensional interface.

9. A system for simulating and predicting the state of energy storage batteries based on digital twins, characterized in that, Specifically, it includes: A multi-scale digital twin model construction module, in which: The current-voltage dynamic response, heat conduction and heat dissipation processes, capacity decay, and internal resistance growth of the battery are modeled using electrical, thermal, and aging sub-models, respectively. A dual-frequency strategy and extended Kalman filtering are combined to dynamically update the model parameters. Adaptive gating parameters are introduced to dynamically fuse the outputs of the physical and data-driven models, achieving high-precision state estimation through dynamic weighted fusion and residual compensation mechanisms. A multi-level structure mapping engine is constructed to achieve bidirectional mapping between different systems: cell, PACK, cluster, and module. This multi-level structure mapping engine uses the cell-level fused state as the basic input and establishes a continuous correlation from the micro-cell state to the macro-system performance by introducing electrical topology constraints, thermal coupling relationships, and physical attribute weights. The multi-level structure mapping engine employs a bidirectional mapping mechanism of uplink aggregation and downlink reconstruction. Uplink aggregation is used to progressively converge the cell-level state to the PACK, cluster, and module levels. Downlink reconstruction is used to reverse-calculate the cell-level state based on historical mapping relationships and weight distributions for system-level analysis or anomaly localization requirements. The trend simulation and state prediction module contains: A coupling mechanism is constructed using electrical sub-models, thermal sub-models, and aging sub-models. Cross-scale state convergence and feature reconstruction are performed by combining multi-temporal residual learning with physical baselines. A multi-temporal anomaly judgment system is constructed to calculate the physical and data residuals of key monitoring indicators. Combined with rate of change detection, time-series anomaly detection models, and long-term trend analysis, anomaly localization is performed through an electrical-thermal-aging collaborative feedback mechanism. By quantifying feature contributions, counterfactual simulation, and feature attribution analysis, combined with attention mechanisms and causal analysis, multi-level explanatory information is generated. The visualization and intelligent early warning module includes: A three-dimensional spatial topology model of cell-PACK-cluster-module is established. The cell status is aggregated to the PACK level, cluster level and module level through hierarchical mapping rules to maintain the topological relationship in physical space. Key state parameters are dynamically bound to visual primitives, and the values ​​are mapped to visual attributes through normalization, weighting and topological constraints. Time-series recording and visualization of key abnormal indicators; Alarm signals are generated based on trend prediction results and abnormal trajectories and are processed in a hierarchical manner, employing adaptive threshold calculation, multidimensional risk scoring, and historical pattern comparison.

10. The system according to claim 9, characterized in that, The system employs a cloud-edge-device architecture for power plant system state simulation and prediction, and the specific implementation method is as follows: The multi-scale digital twin model construction module is deployed on the edge side to collect and preprocess current, voltage, temperature and initial state parameters in real time. Based on extended Kalman filtering, high-frequency state estimation and low-frequency parameter adaptive update are performed on cell SOC, SOH, polarization voltage and temperature to generate cell-level fine state information. The nonlinear features captured by the data-driven model are supplemented by a gated fusion strategy to obtain the instantaneous state. The trend simulation and state prediction module is deployed on the cloud side to receive instantaneous state data from the edge side. It uses thermal-electric-aging joint modeling and time series prediction network to predict the short-term, medium-term and long-term battery operation trends, degradation trajectories and risk indicators. It also combines confidence calculation and feature contribution analysis to generate adaptive alarm signals based on the prediction results and provides graded early warning and operation and maintenance suggestions in combination with multi-dimensional risk scoring. The visualization and intelligent early warning module is integrated into the system. Through the state-structure binding mechanism, multi-level simulation data is mapped to the topology model of the compartment / cluster / PACK / cell. The battery status, operating trend and abnormal evolution process are displayed intuitively through dynamic colors, deformation and animation.

Images